A comprehensive framework for assessing marine, nearshore and transitional waters quality integrating Irish Water quality Index (IEWQI) model from remote sensing products using computational intelligence techniques

The research focused on developing the framework for assessing marine, nearshore and transitional waters across Ireland and validated for generalization of the framework across at any geospatial scale using remote sensing (RS) products. To the best of authors knowledge, existing most of the studies only have demonstrated for retrieving particular water quality (WQ) indicators like turbidity, salinity or chlorophyll a without in depth validation results. Recently the authors comprehensively reviewed several studies focusing on the RS applications for assessing WQ using computational intelligence techniques (CIT) like machine learning, artificial intelligence, statistical approaches etc. Unfortunately, the reviewed findings reveals that most of the research are questionable in terms of using data transparency, and validation with independent or other geospatial domains applications of the existing developed tools. Therefore, the research aim was to develop a novel framework and validated with independent datasets including new domain(s) adaptation or validation. For developing the framework, to achieve the goal of the research, the study utilized Sentinel-3 (S3) OLCI RS reflectance data. For obtaining RS data, the study utilized S3-OLCI level 3(L3) and level 4 (L4) reflectance data Rhow_1 to Rhow_11 form the Copernicus Marine Services (CMS) repository datasets for 2016 to 2024. To obtain the overall WQ, the research considered 49 (in-situ) EPA, Ireland monitoring sites across various transitional and coastal waterbodies for computing the overall WQ (IEWQI scores) scores using recently developed and widely validated the IEWQI model. After than the RS data prepared and match-up with 49 considering monitoring sites. For predicting IEWQI scores, the research utilized the multi-scale signal processing framework (MSSPF) by following configurations: data augmentations: 2x to 20x, noise level from 0.0001 to 0.05, and data spilled ratios 60-20-20 and 70-20-10, respectively for train, test and validation of 43 CIT models using RS data from 2016 to 2023 both L3 and L4, whereas the 2024 dataset using for testing independent dataset to generalize the model prediction capabilities. Utilizing four identical model performance evaluation metrics, the results reveals that the PyTorchMLP could be effective (train performance : R² = 0.86, RMSE =0.09, MSE = 0.008, and MAE = 0.067; test performance : R² = 0.84, RMSE =0.094, MSE = 0.008, and MAE = 0.071; and validation performance : R² = 0.81, RMSE =0.095, MSE = 0.009, and MAE = 0.074, respectively at 7x augmentation with 0.0001 of noise level for 60-20-20) compared to the 43 CIT models in terms of predicting and validating independent dataset (independent dataset validation performance for 2024 : R² = 0.62, RMSE =0.164, MSE = 0.026, and MAE = 0.12). Based on the predicted IEWQI scores, the WQ ranked “marginal”, “fair” and “good” categories for Irish waterbodies. The findings of the framework align with the traditional EPA, Ireland monitoring approaches. However, findings of the research reveals that the proposed framework could be effective to monitoring WQ general purposes using RS data across any geospatial resolution.

Keywords: remote sensing; Copernicus database; MSSPF, IEWQI, Ireland.